LinkedIn
Most accounts I inherited were not stuck in learning. They were stuck in editing: a new conversion goal on Monday, a budget jump on Wednesday, a campaign pause on Friday, then a complaint that Smart Bidding never settled. The useful question is not just how long learning lasts. It is what keeps changing underneath the bids.
This list is for anyone running Smart Bidding who sees Learning or Learning (limited) more often than they would like. In Google’s updated guidance, the visible Learning label lasts up to seven days and has three stated triggers: a new or reactivated strategy, a change in strategy composition, or a significant conversion-goal change. Other edits can still move performance without showing that label. Google says ads keep serving and historical data still counts. Neither promise means you can change the inputs every other day and expect steady results.
1. Fix broken conversion tracking before changing bids
A dead tag, a consent banner that blocks the purchase event, a GA4 import that stops syncing: each can make a working account look broken. Smart Bidding relies on conversion data. When that feed fails, the system cannot make the same decisions from the same information; fixing the feed can leave you with another period of noisy results as new conversions arrive. Check the signal before you blame the target.
I would start in Conversions > Diagnostics and check the GA4 import status before touching tCPA or tROAS. If tracking is broken, repair it and give the account a full conversion cycle before judging the bids. Google’s learning guidance makes clear that historical and new data both matter. A clean target cannot compensate for a missing purchase event.
2. Freeze conversion goals while you judge performance
Add a booked-call action alongside a form fill, or remove a micro-conversion, and you may change what the strategy is trying to win. A significant conversion-goal change is one of Google’s stated Learning triggers. The mechanism is plain: yesterday’s bids were trained against yesterday’s definition of success. Change that definition and the old results become a less useful guide.
Make primary goals deliberate, then leave them alone during calibration. If you need to observe a new action without asking Smart Bidding to optimize for it, set it up as secondary first. When a primary-goal change is necessary, make it in one window rather than alongside a new target and a campaign restructure. Then judge the result over two conversion cycles or two weeks, whichever is longer. The calibration recap is worth reading before you change a Search, Shopping, or Performance Max goal.
3. Switch bid strategy only when you mean to change the job
Moving from Max Clicks to tCPA, or from tCPA to tROAS, changes what the auction model is asked to predict. I used to describe a strategy switch as plumbing. I was wrong. The account can keep its history, but past bids do not answer the new question in quite the same way. Google lists a new or reactivated strategy as a reason for the Learning label.
Switch once, then hold the other inputs steady. Keep campaigns and conversion goals fixed while you evaluate the change over two cycles or two weeks, whichever is longer. If you are still deciding what the dropdown should say, make that decision first; I lay out the tradeoffs in this bidding strategies guide.
4. Lock value, counting, and attribution settings during a test
Editing value rules, changing Every to One counting, or changing attribution can make the same traffic produce different reported results. A tROAS strategy then sees a different value per click, even if the customer and the order have not changed. This is why I treat signal quality as seriously as strategy choice. Do not rewrite the scorecard halfway through the match.
First, record the existing value per conversion and settings. If a change is necessary, document the before and after, make the change without a simultaneous target adjustment, and wait two full cycles before judging tROAS. A tidy-looking settings edit can have an untidy effect on the data bids use.
5. Consolidate campaigns in one planned change window
Consolidation can help an account. Doing it piecemeal while you assess performance does not. Moving keywords, splitting a campaign, or merging campaigns changes where traffic sits in relation to a bidding strategy. Google names a change in strategy composition as a Learning trigger; practitioner reset tables flag restructures, too. History remains, but the new structure asks the model to use it differently.
Tidy once, then freeze. Batch the moves, leave goals and targets untouched in that window, and save the old structure. Your first comparison should be against the outcome you actually sell to the business, not whether the campaign list looks nicer. What the deck calls simplification can be three separate changes wearing one coat.
6. Stage large budget changes instead of making one jump
A budget move does not itself show the Learning label. It still changes how much the strategy can buy. Double a daily budget and the system has to find more eligible auctions; halve it and you restrict the auctions it can enter. Neither move erases history, but both can make yesterday’s CPA a poor forecast for tomorrow. That distinction between a label trigger and a performance wobble matters.
Do not confuse ‘no Learning label’ with ‘no adjustment period.’ If the business does not need the change immediately, move in steps under 30%, wait through a conversion cycle, then assess whether to move again. If it does need the change now, make it and give the results time to settle instead of trying to repair the first uneven day with another budget edit.
7. Move tCPA or tROAS targets once per conversion cycle
A tROAS move from 500% to 300% tells the strategy to pursue a different set of auctions. Google says target changes apply in the next auction without the Learning label, and its example of a move from 500% to 400% still calls for a conversion cycle to settle. The familiar 20% rule is not policy; the practical problem is the size and frequency of the change, especially when you stack moves inside one cycle.
One target move, one observation window. Make the change, write down the date, and judge it over two cycles or two weeks, whichever is longer. Do not nudge it again on Friday because Wednesday’s chart looked rude.
8. Test a new landing page before sending it everything
I have watched a landing-page swap damage an account while the ads and keywords stayed put. The bid strategy was still predicting conversions from the old page’s performance; visitors were now seeing a different page. A page change need not display Learning to change CPA. The Quality Score guide covers the page-experience side, but the immediate bidding issue is simpler: the conversion rate behind the prediction may have moved.
Keep a comparison alive. Put the new page in an experiment split, send it 50% of traffic, and compare CPA over two cycles before moving all traffic. If the page is worse, a bid adjustment is a poor substitute for fixing the page. If it holds up, you can make the change without having mistaken a quiet Learning label for proof that nothing happened.
9. Keep a paused campaign’s restart separate from other edits
A long pause is the only item here where doing nothing becomes the disruptive change. Conversion data stops arriving while the campaign is off, and the auction may not look the same when it returns. A reactivated strategy can also show Learning under Google’s guidance. I used to treat a holiday pause as a clean stop-and-start. It is cleaner for the calendar than for judging CPA.
If you can keep spending, consider a lower budget rather than a full pause. If you must stop, stop. On restart, do not also change the goal, target, and landing page; give the resumed campaign a readable comparison. This practitioner discussion of pauses and stale data makes the same practical point: the return needs attention, even when the pause itself looked harmless.
10. Test broad match or AI Max expansion away from the core
Broad match or AI Max search expansion can change the query mix without adding a Learning label. That is not the same as saying the new queries will perform like the old ones. Exact and phrase traffic gave the strategy one set of intent signals; expansion opens a wider pool, including queries it has less evidence for. New inventory deserves its own test.
First, put broad terms in an experiment or a capped test campaign rather than switching the core campaign mid-evaluation. Review search terms and add clear negatives weekly. Move winners into the core only after two stable cycles. The point is not that expansion is bad. It is that you should be able to tell whether expansion worked before you ask it to carry the whole account.
11. End seasonality adjustments when the event ends
A seasonality adjustment tells Smart Bidding to expect a short-lived change in conversion rate. Leave that expectation in place beyond the sale and the model is bidding against an event that has ended. In a low-volume account, there are fewer conversions to help expose the mismatch quickly. Small accounts can need more patience after a change; they do not need a forgotten promo setting added to the mix.
Set the end date when you create the adjustment. Put a check on the calendar for the morning after the event and confirm the setting has ended. That is the first step, not a retrospective search for why ordinary Tuesdays suddenly look expensive.
12. Make a modest target adjustment without fearing the label
Not every target change is the large jump in entry 7. A move from $50 to $55 tCPA still changes bidding, but it does not create a new strategy or, by itself, show Learning. Google says the new target applies in the next auction. No label does not mean no effect; it means you can make a measured change without treating the status column as a stop sign.
If CPA has run 15% hot for a full cycle, consider a 10% target move toward reality. Log the date and leave it alone for the next cycle. I would rather make one legible adjustment on Tuesday than three tiny ones nobody remembers making by Friday.
13. Add an audience in observation mode to learn, not intervene
An in-market or first-party audience set to Observation lets you examine performance without making it the campaign’s targeting rule. Audience changes do not, by themselves, trigger Learning. The useful distinction is between collecting a view of the traffic and changing which traffic you pursue.
Start with observation and no bid adjustment. Let it collect data for two cycles, then decide whether a creative change or separate test is worth making. Do not quietly switch it to targeting halfway through that window and call the before-and-after report a test.
14. Trim a weak schedule edge, not the whole day
Removing an hour or two that spends but brings in no conversions can cut low-value auctions without changing the goal or strategy composition. The campaign still draws on the rest of its history; ads keep serving while background learning continues. I treat a small schedule edit like maintenance, not surgery. Maintenance still needs evidence.
Check volume before you cut. Pull the last 60 days of hourly performance and look for hours with meaningful spend and no conversions. Exclude only the weak edges, then revisit the result after a cycle. One empty hour in a thin account is not a verdict; it is a small sample wearing a confident expression.
15. Give a responsive search ad fresh assets
New headlines or descriptions change the message, not the conversion goal or strategy composition. Creative edits are outside the bid-strategy restart list. They can still change which ad people click, which is why I would measure them rather than declare them risk-free. But there is no reason to leave stale copy untouched solely because the account is sensitive to Learning.
Keep the current winners running. Add two headlines and one description, then review the assets after 500 impressions before cutting anything. A new line about pricing or same-day quotes gets a chance to prove itself without making the entire account a new bidding experiment.
16. Pause one weak ad without pausing the campaign
Pausing an underperforming responsive search ad is different from stopping its campaign for days. The remaining ads can still serve against the same keywords and conversion goal; the message mix narrows, but the bid strategy has not been replaced. Google says creative changes do not remove history or trigger Learning. Do not keep a weak ad alive to protect a status label.
First, check that the ad really has enough data to call it weak. Pause it while the stronger ads keep running, then add a new RSA with a different hook. I would leave headlines unpinned unless compliance requires otherwise. You are testing a message, not commemorating an old one.
17. Add clear negative keywords before you touch a target
This is the edit I would start with in a messy account. A negative keyword removes auctions you did not want; it does not change the conversion goal or create a new bid strategy. Practitioner reset tables treat negatives as routine maintenance rather than a bidding restart. They can change the traffic mix, of course. That is the point: stop paying for plainly irrelevant queries before asking the target to make irrelevant traffic profitable.
Pull the last 30 days of search terms and exact-negate the clear junk. Batch the less obvious n-gram work weekly, and leave the target alone while you see what the exclusions do. If I had one hour, I would do that first, then check the conversion feed and any expired seasonality adjustment before closing the laptop for a couple of cycles. It is the same discipline behind how groas works: continuous execution, logged changes, and a strategist accountable for the outcome. Not another week of edits nobody counts.